Exploring novel noise reduction methods for Pearson-correlation-based brain graph analysis: a perspective from autism spectrum disorder diagnosis

In recent years, enormous efforts have been made in analyzing brain networks to reveal the correlation between brain connectivity abnormality and neurodevelopmental diseases such as autism spectrum disorder (ASD). This type of research typically uses Pearson correlation coefficients to estimate the connectivity between brain regions and is often limited by noise present in the data. Existing methods generally focus on mitigating task-related noise by selecting the most discriminative features. However, other sources of noise could also have a significant influence and should be taken into account. In this paper, we examine three possible sources of noise: the instability of the Pearson correlation coefficient results from short time series length, the interference between neighbouring brain regions, and the changes in the global brain state. Additionally, we propose three methods to address these three types of noise respectively, namely unreliable feature suppression, neighbour interference elimination, and global signal normalization. In order to validate the reliability of our findings and the effectiveness of our proposed methods, we conducted experiments on the publicly available dataset ABIDE. Our results showed a diagnostic accuracy of 73.22%, with nearly 3% of the accuracy attributed to our proposed methods. Furthermore, based on the denoised features, we identified the brain areas that are significant in distinguishing ASD from healthy controls.